Rosalie Viney and Elizabeth Savage CHERE, University of Technology, Sydney

1

Health care policy evaluation: empirical analysis of the restrictions

implied by Quality Adjusted Life Years

Rosalie Viney and Elizabeth SavageCHERE, University of Technology, Sydney

Preliminary – not for quotation

2

Health care policy evaluation

• Resource allocation decisions in health care increasingly based on cost-effectiveness analysis– Australia

• Pharmaceutical Benefits Advisory Committee (PBAC)

• Medicare Services Advisory Committee (MSAC)

– United Kingdom• National Institute of Clinical Excellence (NICE)

– Canada (Ontario), Denmark, Finland, Netherlands, Sweden• All have formal requirements for economic evaluation for

reimbursement decisions for pharmaceuticals

– United States• Discussion of a PBAC type process for new pharmaceuticals

3

Example: Pharmaceutical Benefits Scheme

• PBAC required by legislation to consider cost-effectiveness (among other factors)

• Company seeking PBS listing for a drug must submit an economic evaluation– Evidence of clinical effectiveness– Estimate potential health gain from new drug relative to current

treatment (measure of effectiveness)– Estimate additional cost of new drug relative to current treatment

• proposed price for new drug

• standardised “prices” for other health services

– Calculate incremental cost-effectiveness ratio

4

Examples of PBAC Decisions

Incremental cost per additional Quality Adjusted Life Year gained “league table”

Submission $/QALY Recommendation

1 $4,690 Recommend 2 $5,244 Recommend 3 $8,570 Recommend 4 $10,530 Recommend 5 $13,121 Recommend 6 $17,937 Not Recommend 7 $21,225 Recommend 8 $24,343 Recommend 9 $133,337 Reject

Source: George B, Harris A and Mitchell A. Cost effectiveness and consistency of decision making: evidence from pharmaceutical reimbursement in Australia (1991-1996). Pharmacoeconomics 19:1103-1109 2001.

5

Measures of health gain (effectiveness)

• Ideally want measures that are comparable across different diseases and treatments

• “Intermediate”/ “surrogate” (clinical) outcome not comparable

• Life years saved

• Quality adjusted life years saved (QALYs)

6

Estimating health gain

• Decision tree framework• Epidemiological information used to identify

– disease states– outcomes of treatment– probability of different outcomes (proportion of a population

experiencing the outcome)– survival (durations) associated with different outcomes– quality of life associated with different outcomes

• Estimate expected life years associated with each alternative (new treatment, current treatment)

• Not all years of life have the same value (quality)

7

Quality Adjusted Life Years

• “Quality Adjustment” reflects valuation of poorer health states relative to “full health”

• QALYs calculated by weighting survival time associated with each treatment outcome by a QALY weight

• QALY weight of 1 for a year of life in “full health”• QALY weights generally lie between 0 (death) and 1 (full

health)– “Worse than death” health states permissible

• QALYs measure strength of preference for survival and quality of life and trade-offs between the two

8

Example: QALY WeightsAIDS, CD4 count range 0-50 0.79

Arthritis, before treatment 0.609

Arthritis, after treatment 0.647

Cancer, breast, after surgery, first recurrence 0.85

Cancer, breast, after surgery, third recurrence 0.3

Cancer, prostate, metastatic, early progressive disease

0.83

Proctitis/cystitis after randiation therapy for prostate cancer

0.9

Stroke, severe, motor deficit 0.03

Stroke, mild cognitive deficit 0.54Source: Tengs, T and Wallace A. One Thousand Health-Related Quality of Life Estimates

Medical Care 38(6) 583-637, 2000

9

• Health care can be characterised as lotteries over profiles of consumption and health

• Treatments modify the set of lotteries

• Individuals will have preference orderings over alternative treatments

• QALY model implies restrictions on form of utility function

mmT

mT

mmTT pchchpchchUU ;,,,,,,;,,,, 11

11111

11

Economic underpinnings of QALYs

10

Outline

• Describe methods used to derive QALYS• Outline the restrictions on preferences imposed by the

QALY model• Describe a discrete choice experiment designed to

investigate preferences for health care• Test restrictions on preferences implied by the QALY

model

11

QALYs

• Measure the value of a particular health state relative to full health, ignoring consumption

• Under the QALY model

where v(ht) is the quality (utility) weight of health state in period t

• ie identical additive separable function in each period

T

ttT hvhhhU

121 ,...,,

12

Quantifying QALY weights

• Involves experiments using standard gambles and time trade-offs

• These rely on restrictions on preferences over consumption and health status over time.

13

QALY model: Standard gamble

• Specify h1

• Find p such that utility of the gamble, B, is equal to A - the certainty of health state h1 for T years

• ie

• Normalise: v(FH) = 1

p

1 - p

Full health, FH for time, T

Immediate death

FHvhvp

ThvTFHvp

1

10.

A

B

h1 for time T

14

QALY model: Time trade-off

• Find x such that utility of the health state 1 for time T, h1 is equal to the utility of full health for a proportion x of T years

ie

FHvhvx

xTFHvThv

1

1 ..

TxT

FH

h1

15

Note:

• Only attributes of the health state and time enter the experiment

• In both cases the experiment sets T

• Individual responses are averaged to determine the QALY weight attached to health state 1

16

QALY model: Application

• Say a treatment improves health status for a proportion, p1, of the population undergoing the treatment from

– h1 (with QALY weight Q1) to

– h2 (with QALY weight Q2) and

– lengthens expected survival time from T1 to T2

• The effectiveness per individual in terms of QALYs gained is equal to p1 (Q2 T2-Q1 T1)

• p1,h1, h2, T1 ,T2 are derived from clinical data

• Q2 and Q1 are derived from the TTO/SG experiment;

– T in the experiment is unrelated to T1 or T2 ;

– p in the experiment unrelated to p1.

17

QALYs: Policy Evaluation

• Effectiveness measured by QALY gain combined with cost to give a C/E ratio.

• If the C/E ratio lies below some threshold level, the intervention is accepted.

• The threshold indicates willingness to pay per QALY.

• Setting the threshold effectively monetises effectiveness.

18

QALY Restrictions

Under what restrictions on preferences is cost-effectiveness analysis based on QALYs likely to lead to health care resource allocation that reflects individuals’ preferences?

19

Literature

• Pliskin, Shephard and Weinstein (1980)• Bleichrodt, Wakker and Johannesson (1997)• Miyamoto, Wakker, Bleichrodt and Peters (1998)

– Preferences defined over a constant health state and survival

• Bleichrodt and Quiggin (1997)– Preferences defined over non-constant health states over time– Expected utility and Rank dependent expected utility

• Bleichrodt and Quiggin (1999)– Preferences defined over non-constant health states and

consumption over time

20

QALY restrictions: B&Q (1999)

• VNM expected utility• Marginality• Symmetry• Standard gamble invariance• Zero condition

21

QALY restrictions: marginality

• Let yG= ($20, H1); yB= ($5, H2); H1>H2

• With marginality an individual is indifferent between– G1: (yB, yB) with prob=1/2 and (yG , yG ) with prob=1/2

– G2: (yB, yG ) with prob=1/2 and (yG, yB ) with prob=1/2

• Both G1 and G2 have the same probability of yG in both periods and of yB in both periods.

• Marginality excludes all complementarity across time periods, such as a dislike for variation or a desire to avoid a bad outcome in both periods.

22

QALY restrictions: symmetry

• Requires that resequencing does not change utility:

• Inconsistent with positive time preference.

T

t

tT yUyyyU1

21 ,...,

23

QALY restrictions: standard gamble invariance

• For all levels of consumption, c and c’

• Utility derived from the health state unaffected by the level of consumption

phcphchc

phcphchc

1,,;,,,

1,,;,,,

24

QALY restrictions: zero condition

• Less controversial

0,cU,cU deathdeath

25

Under QALY restrictions

• z(c) a constant• i.e. Implies constant consumption over time

t

T

t

m

ii

mmT

mT

mmTT

hvczp

pchchpchchUU

11

111111

111 ;,,,,,,;,,,,

26

Implications of QALY restrictions EU: linear in probabilities

• Evaluation function for gambles is linear in probabilities.• No complementarity across time periods.• The utility function over sequences of consumption and

health state is additively decomposable over time • One-period utility functions are the same.

RDEU: probability weighting function• Changes the marginality condition to generalised utility

independence – less strict.

27

Discrete choice experiments and health care

• DCEs widely used in transport economics, environmental economics and marketing

• Used in health economics to:– Explore trade-offs between attributes of health care interventions– Provide monetary valuations of health care interventions– Predict uptake of new programs

• Discrete choice experiments can provide data to explore the underlying utility function for health and health care, particularly trade-off between quality of life and survival (QALYs)

28

DCE Methods

• Stated preference surveys• Choose preferred alternative from a series of

hypothetical choice sets• Alternatives described in terms of attributes• Attributes varied over plausible range of levels of

analytical interest• Experimental design principles used to choose a sample

of choice sets to allow for efficient estimation of parameters of interest

29

DCE Methods: Binary Choice

ikij

ij

ikij

ij

ee

e

ee

e

VVjP

VU

UUjP

VV

V

ikikijiji

ijijij

ikiji

xx

x

''

'

Pr

)Pr(

30

Discrete choice experiment to investigate health care preferences

• Choice between two alternatives: treatment (TR) and non-treatment (NT) for a hypothetical health condition.

PTR

1 - PTR

FH,T ,C

Death

PNT

1 - PNT

H,T

Death

TR NT

31

Attribute levels

• Each alternative has three, four-level, alternative-specific attributes.• Cost of “non treatment” is zero• Health state from treatment is full health or death

PTRProbability of survival

with treatment

TTRLife

expectancy if survive

PRICEPrice

of treat-ment

PNTProbability of survival

without treatment

TNTLife expectancy

without treatment

HSHealth state

without treatment

54% 10 years $20,000 39% 10 years HS1

69% 20 years $40,000 59% 20 years HS2

84% 30 years $60,000 79% 30 years HS3

99% 40 years $80,000 99% 40 years HS4

32

Health states

• Generic health states from EQ-5D multi-attribute utility instrument

• Selected to be ordered, cover the range of health states and to “make sense”

HS1 HS2 HS3 HS4 FHMobility 3 2 2 2 1Self care 3 2 1 1 1Usual Activities 3 3 2 1 1Pain/Discomfort 3 2 2 2 1Anxiety/Depression 3 3 2 1 1

33

You have a health condition that affects your quality of life and your life expectancy. There is a treatment available for your condition and you must decide whether to have treatment or not.

If you do have treatment If you do not have treatment

The out of pocket cost to you is $20,000 There is no out of pocket cost to you

The chances of having successful treatment The chance of surviving or dying with thisor dying are: condition are:

If treatment is successful your life expectancy is If you survive your life expectancy is30 years 20 years

Successful treatment will give you full health Your health is affected by this conditionwhich means you have: in the following ways:No problems in walking about Some problems in walking aboutNo problems with self care No problems with self careNo problems peforming usual activities Some problems in performing usual activitiesNo pain or discomfort Moderate pain or discomfortNot anxious or depressed Moderately anxious or depressed

Would you have treatment?

Yes No

Chance of dying within one month

21%

Chance of surviving

79%

20 years

Chance of dying as a result of treatment16%

Chance of successful treatment84%

30 years

34

Survey design and data collection

• Six 4-level attributes = 4096 choice sets, 64 for each alternative

• 192 choice sets selected from the full factorial• Blocked into 12 separate questionnaires, each with 16

choice sets• Random sample aged 18-60 from Sydney metropolitan

area• door-to-door recruitment; self-completion questionnaires• 347 respondents in total

35

Data

• Treatment was the preferred alternative in 65.1% of observations

• Two choice sets for which 100% of respondents chose treatment

• 9.6% of respondents always chose treatment• 3.7% of respondents always chose no treatment• Remaining respondents responsive to changing

alternatives

36

Methods: Testing QALY modelQALY restricted specification

• Expected utility is given by the sum of the product of probability, time and QALY weight.

• In log form and imposing the QALY restrictions gives

• For rank dependent expected utility P is replaced by a probability weighting function ; the simplest specification is

NT

TRj wherelnlnln jjjjjjj CHTPU

Pg

PPg lnln

37

Methods: Testing QALY modelUnrestricted specification

• With the zero condition the general form is

with

NT

TRj where,,lnln jjjjjjj CTHPgU

jk

jkTjHjjCk

jkjHjjTjj THCHTkk

lnln.

22ln jCjjjCT

KjjkCjH CTCCH

k

38

Notes

• Health states under non-treatment are ordered and dummy-effects-coded in the estimation.

• Loss of consumption is proxied by cost of treatment.• Constant is zero for non-treatment.• Cost is zero for non-treatment.• The impact of full health with treatment is in the

constant. • Expansion is up to quadratic terms.

39

Logit resultsModel 1

(unrestricted)Model 2 (QALY)

Variable Coef P>|z| Coeff P>|z|

Cons -0.3292 0.701 0.941 0.000

LnPtr 1.6919 0.000 - -

LnTtr 0.7157 0.000 - -

Cost -0.0104 0.362 -0.008 0.000

LnTtr_Cost -0.0018 0.500 - -

Costsq 0.0001 0.310 - -

LnPnt 1.4504 0.000 - -

LnTnt 0.5933 0.000 - -

HS1 -0.2192 0.537 -0.853 0.000

HS2 0.4063 0.210 -0.168 0.001

HS3 -0.4169 0.185 -0.386 0.000

LnTnt_HS1 -0.2050 0.068 - -

LnTnt_HS2 -0.1841 0.075 - -

LnTnt_HS3 0.2573 0.010 - -

LnProb - - 1.501 0.000

LnTime - - 0.610 0.000

Log likelihood -3226.699 -3233.59

40

Tests of QALY model restrictions

• Equal coefficients on common TR and NT variables• No interaction effects• Jointly test zero coefficients on interaction variables• Coefficients on probability and time variables = 1

41

Test of equal coefficients on common TR and NT variables

( 1) Ptr= Pnt

chi2(1) = 2.36

Prob > chi2 = 0.1241 CANNOT REJECT

( 2) Ttr= Tnt

chi2( 1) = 0.62

Prob > chi2 = 0.4300 CANNOT REJECT

42

LR test on interaction effects

H0: No interaction effects

Full model Log L = -3226.6992

Constrained model Log L = -3232.6693

LR test:

REJECT

11.94022

11.0705205.0,crit

43

Test for zero interactions in constrained model

• Coefficients on interaction terms = 0– lnrttr_rcost– rcostsq– lnrtnt_hsfx1– lnrtnt_hsfx2– lnrtnt_hsfx3

• chi2( 5) = 10.83

Prob > chi2 = 0.0549 REJECT (JUST)

44

Test: utility function linear in probability and time

( 1) ln(Prob) = 1

chi2( 1) = 39.56

Prob > chi2 = 0.0000 REJECT

( 2) ln(Time) = 1

chi2( 1) = 46.95

Prob > chi2 = 0.0000 REJECT

45

Conclusions• Empirical analysis of this DCE does not provide support

for the QALY restrictions.• Suggests a more general evaluation function is required

for resource allocation in health.

Why the dominance of QALY model?• Monetary measures of welfare gain seen as ethically

troubling.• B&Q: QALY model does not avoid ethical judgements.• Explicit treatment of distribution is preferable.